Hi FlashAttention team 👋
After a long compile session, I successfully built FlashAttention 2.8.0.post2
on an NVIDIA L40S (Ada Lovelace) GPU using CUDA 12.1, PyTorch 2.9, and Python 3.10.
To help others avoid the long build time, I'm sharing the working prebuilt wheel:
✅ Tested environment:
- GPU: NVIDIA L40S (Compute Capability 8.9 / sm_89)
- CUDA: 12.1
- PyTorch: 2.9.0
- Python: 3.10
- OS: Ubuntu 22.04 (Google Cloud VM)
- FlashAttention version: 2.8.0.post2
- Wheel size: ~111.5 MiB
✅ Installation steps:
pip install flash_attn-2.8.0.post2-cp310-cp310-linux_x86_64.whl
pip install --upgrade torch torchaudio torchvision # Recommended
📦 Prebuilt wheel repository:
👉 [https://github.com/h1312200313122003-code/flash-attn-prebuilt-L40S/releases/tag/v2.8.0.post2-cu121-l40s]
Everything works great — verified import, GPU detection, and performance.
Thank you for this amazing library ❤️
Hopefully this wheel saves others a few hours of compile time!
Hi FlashAttention team 👋
After a long compile session, I successfully built FlashAttention 2.8.0.post2
on an NVIDIA L40S (Ada Lovelace) GPU using CUDA 12.1, PyTorch 2.9, and Python 3.10.
To help others avoid the long build time, I'm sharing the working prebuilt wheel:
✅ Tested environment:
✅ Installation steps: